EXECUTING NESTED PARALLEL LOOPS ON SHARED - MEMORYMULTIPROCESSORSSadun

Cache-coherent, bus-based shared-memory multiprocessors are a cost-eeective platform for parallel processing. In scientiic parallel applications, most of the computation involves processing of large multidimensional data structures which results in a high degree of data parallelism. This parallelism can be exploited in the form of nested parallel loops. Most existing shared memory multiprocessors exploit this multi-level parallelism at only one level. In this paper, we explore eecient algorithms and models for executing nested parallel loops and present a simulation based performance comparison of diierent techniques using real application traces. We show that it is possible to exploit the parallelism in nested parallel loops with the use of good scheduling and synchronization algorithms.